Guest Editors Chi Lin and Chang Wu Yu introduce the Special Section on Frontiers in Image and Signal Processing for Remote Sensing.

Interpreting the signal deriving from interferometric synthetic aperture radar (InSAR) analyses in volcanic islands, characterized by strong regional deformations and recurrent seismicity, is a complex and challenging issue. In these zones, the secondary effects connected to the SAR acquisition system cannot be neglected, and it is important to consider that delay phenomena of the electromagnetic waves, due to the propagation in the tropospheric layer and loss of SAR coherence because of dense vegetation, could affect the interferometric phase. This work focuses on Pico, the second largest and the youngest island of the Azores Archipelago (North Atlantic Ocean). This island consists of a central volcano and a fissure zone. These systems are inactive but recurrent microseismicity occurs in a rock volume hosting a partially crystallized magma storage system, which fed the recent activity of the central volcano. In the same area affected by microseismicity, the main volcanic edifice shows flank instability. All these elements support the hypothesis of possible reactivation of the shallow magmatic system. Aiming to check potential active ground displacements and to define their source, we collected two datasets of C-band Sentinel-1 SAR data, both in descending and ascending acquisition geometry, from January, 2017, to December, 2020. The application of the small baseline subset method of differential InSAR allowed drawing the mean ground velocity maps over the island and the displacement time series, useful to understand the deformation evolution. InSAR data only evidence areas affected by small-scale subsidence at the cinder cones of the fissure zone and along the southeastern slope of Pico volcano, where local debris flows activate during rainy periods.

Remote sensing is an effective way to analyze land surface changes on regular basis globally. In the previous literature, numerous change detection models were developed to detect the multitemporal or seasonal variations using different optical datasets, such as multispectral and hyperspectral. But there are many challenges involved in the processing of numerous bands, especially in the case of hyperspectral datasets, such as computational constraints and radiometric/atmospheric distortion. A simple framework-based artificial neural network (ANN) and post-classification comparison (PCC), named ANN-based PCC (ANPC), has been proposed to detect the multitemporal changes over agricultural land using hyperspectral, i.e., Earth Observation (EO-1) Hyperion and multispectral, i.e., Landsat-8. For cross-referencing, a comparative analysis is performed with respect to well-known PCC-based change detection methods, i.e., random forest-based post-comparison (RFPC) and support vector machine-based post-comparison (SVMPC). Experimental outcomes confirmed the effectiveness of ANPC (with an accuracy of more than 90%) in the extraction of multitemporal changes as compared to RFPC and SVMPC (with an accuracy of <90 % ). Our study enhances the utilization of the hyperspectral dataset (due to narrow spectral bands) in the extraction of critical information about the earth’s surface parameters.

Remote sensing image retrieval (RSIR) frameworks encounter several issues in practical scenarios with (1) the dominating repetitive features in the final image representation, (2) inter- and intra‐class variability across objects, and (3) time complexity in exhaustively searching the large-scale remote sensing image archives. Motivated by these facts, we propose a deep feature-splitting approach that enhances a localized hashing (DFS-LHash) model for RSIR. The DFS strategy splits the fully connected (FC) layer features into equally sized blocks for a split-based localized hash learning, keeping VGG-16 as the baseline network. We incorporate effective deep clustering to improve classification performance by reducing the center-cluster loss. A block-wise softmax is applied on each block to pay more attention to the essential features with reduced dimension for object localization, which enhances retrieval performance. The center-cluster loss applied strengthens the class discriminant information and minimizes the distance between the feature descriptors belonging to similar classes to a greater extent. The classification error from the DFS strategy is reduced to learn highly discriminant and localized binary codes effectively. The proposed model provides effective search with better retrieval time and achieves state-of-the-art performance with mean average precision values of 97.42%, 96.06%, and 93.47% for the University of California- Merced, PatternNet and aerial image datasets, respectively.

Generalized zero-shot learning (GZSL) is the most popular approach for developing ZSL, which involves both seen and unseen classes in the classification process. Many of the existing GZSL approaches for scene classification in remote sensing images use word embeddings that do not effectively describe unseen categories. We explore word embedding to describe the classes of remote sensing scenes to improve the classification accuracy of unseen categories. The proposed method uses a data2vec embedding based on self-supervised learning to obtain a continuous and contextualized latent representation. This representation leverages two advantages of the standard transformer architecture. First, targets are not predefined as visual tokens. Second, latent representations preserve contextual information. We conducted experiments on three benchmark scene classification datasets of remote sensing images. The proposed approach demonstrates its efficacy over the existing GZSL approaches.

High-resolution remote sensing (HRS) images are widely used in regional planning and mapping, and the timely update of building change information is essential. Although HRS images offer rich surface information for building change detection, the complexity of the building background and its varying nature require the network to capture both local and global information. To address this issue, the efficient-UNet++ change detection network is designed with UNet++ as the baseline model and EfficientNet-B1 as the backbone network to retain more semantic information. To extract features efficiently and reduce the network parameters, some standard convolutions are replaced with depthwise separable convolution (DW Conv). The network is also equipped with multi-headed self-attention to improve its ability to capture global information. The proposed methods achieve an intersection over union of 0.7921 and an F1 score of 0.8804 on the LEVIR-CD dataset.

We are focused on improving the resolution of images of moving targets in inverse synthetic aperture radar (ISAR) imaging. This can be achieved by recovering the scattering points of a target that have stronger reflections than other target points, resulting in a higher radar cross section of a target. However, these points are sparse and moving targets cannot be correctly detected in ISAR images. To increase the resolution in ISAR imaging, we propose the fast reweighted trace minimization (FRWTM) method to retrieve frequencies of sparse scattering points in both range and azimuth directions. This method is a two-dimensional gridless super-resolution method that does not depend on fitting the scattering point on the grids. Using computer simulations, the proposed algorithm is compared with fast reweighted atomic norm minimization (FRANM), sparse Bayesian learning (SBL), and SL₀ algorithms in terms of mean squared error (MSE). The results show that FRWTM performs better than the other methods, especially SBL and SL₀ at low signal-to-noise ratio (SNR) and fewer samples.

We propose a modification of a water balance method (NDVI-Cws) aimed at improving the estimation of actual evapotranspiration in forest areas. The improvement concerns the calibration of the Cws meteorological factor, which regulates the sensitivity of forest ecosystems to short-term water stress. Such calibration is based on Satellite Application Facility on Land Surface Analysis evapotranspiration products, which are informative on the actual impact of water stress on plant evapotranspiration. The original and calibrated versions of NDVI-Cws are applied in two environmentally diversified forest areas in Central Italy, where water fluxes were measured by the eddy covariance technique. The first site corresponds to a Mediterranean coastal pine forest (San Rossore) and the second to a mountain beech forest (Collelongo), where water fluxes were measured for the years 2001 to 2005 and 2000 to 2009, respectively. The calibration performed leaves unchanged the model setting for San Rossore while it induces a reduction of the model sensitivity to water stress for Collelongo. The calibrated NDVI-Cws version yields optimal evapotranspiration estimation accuracies for both study sites; the determination coefficient is around 0.90, the root mean square error is lower than 0.30 mm day ^{− 1} and the mean bias error is around ±0.01 mm day ^{− 1}. These findings indicate that the modified water stress factor is a realistic descriptor of the soil–vegetation–atmosphere response of forest ecosystems to water shortage.

Land surface temperature (LST) is an important climate variable used to assess the effects of climate change. This research project aims to compare the results of mono-window (MW) and split-window (SW) algorithms against the Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat 8 Collection 2 Level-2 surface temperature (L8-C2L2) products and identify the most suitable techniques. In-situ measurements of the earth’s surface temperature and relative humidity have been recorded in 2020. LSTs have been validated using root mean square error (RMSE) based on the in-situ meteorological data. Validation analysis has indicated that the SW algorithm combined with in-situ micro-scale atmospheric water vapor content values was more accurate for LST. In the overall study area, the RMSE values of 1.09°C, 3.97°C, 4.36°C, and 6.80°C have been calculated for SW, MODIS, L8-C2L2, and MW LSTs, respectively. These results have demonstrated that the SW algorithm outperformed the other LST products. The maximum difference between the in-situ earth surface temperature and the SW algorithm was 0.79°C. These findings are essential for comparing different data-driven approaches and identifying the most efficient techniques. The study’s significance lies in identifying the most appropriate method for LST retrieval, which can aid in climate change studies and inform decision-making processes.

Individual tree segmentation of forestry point cloud data is of great significance to forest management and resource detection, because it can quickly and efficiently extract tree parameters and calculate biomass. Although research on individual tree segmentation of forestry point cloud data has made great progress, there are still many problems. For example, it is difficult to separate two trees when they are close to each other or occluded. In this work, a point cloud segmentation method is proposed to obtain an individual tree from forest plantation datasets, which combines improved point transformer and hierarchical clustering method. First, we use the improved point transformer to remove the ground and non-tree data to obtain pure tree point cloud data. Second, the tree point cloud data are converted to digital surface model and the watershed segmentation algorithm is used for the preliminary segmentation. Subsequently, a merging algorithm is proposed to merge the missing segmented point cloud data at the edge of the point cloud with the successfully segmented point cloud data, according to the nearest point cloud category. However, the results after the merging algorithm still have trees that cannot be segmented. Finally, a hierarchical clustering method is proposed for fine segmentation. For the improved point transformer, we utilized three regions for verification and three regions for testing. The mean intersection over union (MIOU) of the improved point transformer on the test set is 0.976, which is 1.1% higher than that of the original point transformer. For individual tree segmentation, we tested on five regions and obtained a MIOU of 0.742. The results demonstrate that the method proposed in this work can achieve better individual tree segmentation than other methods.

National Oceanic and Atmospheric Administration (NOAA)-20 (formerly Joint Polar Satellite System-1) is a polar-orbiting weather satellite launched on November 18, 2017. Visible-Infrared Imaging Radiometer Suite (VIIRS) is one of five instruments onboard NOAA-20 (N20), and it jointed on orbit the previous VIIRS instrument operating on the Suomi NPP (SNPP) spacecraft since November 2011. During post-launch testing, the accuracy of VIIRS sensor data record (SDR) geolocation products was evaluated using a system called ground control point matching. The control points are based on 30-m Landsat images as “ground truth,” and the matchups employ their cross-correlation with the 375-m VIIRS images in band I1. Initial N20 VIIRS geolocation errors shortly after launch exceeded 2 km and were somewhat larger than those previously observed for SNPP VIIRS. The optimization of processing parameters successfully reduced the geolocation uncertainty to less than 400 m (at nadir, 3-sigma), which is comparable to the current performance for the SNPP VIIRS SDR products, with the majority of the geolocation errors within 200 m. The improvement was achieved by adjusting only the instrument-to-spacecraft rotation angles (the “mounting matrix”). Although the uncertainty evaluation was based on global matchups collected over 12-month periods, the parameter optimization was based on a reduction of the geolocation angular errors by reprocessing VIIRS SDR and re-analyzing the global matchups from 16-day orbit repeat periods. Geolocation of the 750-m day/night band (DNB), which is not co-registered with the 375-m I-bands and the 750-m M-bands, was evaluated and optimized separately, using the correlation between DNB and the I1 band re-projected images. Initial N20 VIIRS DNB geolocation errors were even larger than those observed for the other bands, but they were also successfully reduced by adjusting DNB-specific processing parameters. The initial improvements in the N20 VIIRS geolocation accuracy were achieved in 2018, and the processing parameters were further refined after an optical model correction in 2021.

The high-precision and high spatio-temporal resolution settlement time series data generated by integrating Global Navigation Satellite System (GNSS) and Interferometric Synthetic-Aperture Radar (InSAR) data are of great value for the safe operation of high-speed railways. The GNSS monitoring stations and InSAR monitoring area present zonal distribution along the high-speed railways, which causes the spatial matrix of the GNSS and InSAR data fusion model to be an ill-conditioned matrix. The inverse of the ill-conditioned space matrix is unstable, and it is difficult to estimate the optimal parameter value of the fusion model, which leads to low accuracy and large fluctuations in the fusion of GNSS and InSAR data. We propose a spatio-temporal filter fusion (STFF) method to solve the influence of the ill-conditioned space matrix on the fusion of GNSS and InSAR data, which realizes the high-precision fusion of GNSS and InSAR data along the high-speed railway. First, we propose the kriging model of the adaptive spectral correction method to construct the spatial model of GNSS and InSAR data, avoiding the generation of the ill-conditioned space matrix. Second, iterative almost unbiased estimation is used to obtain the weights of GNSS and InSAR data, avoiding the occurrence of negative variance components. The experimental data are fused using STFF, the spatio-temporal Kalman filter (STKF) method, and the spatio-temporal random effect (STRE) fusion method, respectively. The research results show that the comprehensive performance of STFF is significantly better than that of STKF and STRE. In the spatial domain, the root mean square error (RMSE) reduced by 47% and 46% compared with STKF and STRE, respectively, and the structural similarity index increased by 22% and 16% compared with STKF and STRE, respectively. In the time domain, the RMSE reduced by 29% and 34% compared with STKF and STRE, respectively.

Benefiting from the advantage of weather and illumination independence, less processing data and shorter data time span, spaceborne interferometric synthetic aperture radar technology has certain potential in building 3D reconstruction and urban DSM reconstruction because of its sensitivity to height. However, the phase mixture in the building layover area makes it difficult to realize the accurate building height reconstruction. Therefore, an interferometry model of high-rise buildings was established by comprehensively analyzing the relationship between building structure, scattering mechanisms, and the geometry of interferometric imaging. At the same time, a model-based height estimation method was proposed for high-rise buildings with complex structures and mixed layover interferometric phase. The established interferometry models of buildings are the basis of the height estimation, which reveals the phase mixing mechanism, phase variation characteristics, and the relationship between phase and height. Further, based on the established model, the characteristics and difference of the mixed layover interferometric phase and the pure facade interferometric phase can be obtained. Then, based on these characteristics and by detection of the facade-only imaging area, the pure facade interferometric phase can be recovered from the mixed layover interferometric phase, and the building facade height can be reconstructed by the conversion between phase and height. A pair of TerraSAR-X data located in Harbin city, China was used for experiment and the results show that the height estimation algorithm proposed in this paper can effectively obtain the building height and the DSM in urban area, and for the results of building heights, the average absolute error is within 2 m and the root mean square error is between 2 and 4 m.

Accurate semantic segmentation of images has long been a research priority in remote sensing. However, the presence of geometrically complex and spatially diverse objects increases the difficulty in simultaneously obtaining coherent and accurate labeling result. To solve this challenge, our study combined multiscale geometric feature extraction with convolutional neural network and proposed a new U-shaped contourlet network (USCNet) for segmentation from high-spatial-resolution remote sensing images (HSRRSIs). This network is designed to learn and characterize the geometric features present in HSRRSIs. The USCNet first transforms the original dataset into a pyramidal structure containing multiscale and multidirectional geometric information and then fuses the spatial and geometric features to extract high-level semantic information. This network has two advantages: (1) coarse-to-fine spatial features are learned efficiently using a hierarchical learning structure and (2) the multiscale learning scheme captures geometric information in different directions. The results of extensive experiments conducted on two remote sensing datasets (the International Society for Photogrammetry and Remote Sensing Vaihingen and Potsdam challenge datasets) show that the proposed approach outperforms several state-of-the-art semantic segmentation methods.

Regional hydrological analysis generally requires meteorological inputs with adequate spatial resolution and coverage. Satellite-derived precipitation covers relatively large areas at various temporal scales. The global precipitation measurement (GPM) began releasing a new generation of global precipitation products in April, 2014, i.e., the integrated multi-satellite retrievals for GPM (IMERG), which has a spatial resolution of 0.1 deg (latitude) × 0.1 deg (longitude). However, IMERG does not have the sufficient resolution required by the most advanced fine-scale hydrological models. Meanwhile, due to the randomness of daily precipitation, it is difficult to obtain stable precipitation influence factors. The results of this work show that: (1) The IMERG V06 shows obvious multifractal characteristics, which makes it possible to use the multifractal method to improve its spatial resolution without the help of other elements. However, the accuracy of precipitation products will suffer a certain loss. (2) The downscaling method considering the influence factors of precipitation did well on a monthly dataset, and the maximum CC can reach 0.911. At the same time, the random forest regression model is significantly better than the traditional multiple linear regression model since the former is better matched with the original monthly precipitation data and can produce more local details. (3) The downscaled monthly precipitation data are helpful to the spatial heterogeneity recovery of daily precipitation. The recovery can enrich the spatial details of the daily precipitation and improve the accuracy to a certain extent. Compared with the multifractal (MF) downscaling results, the accuracy of the MF-RFR model was improved by 10.3%, whereas the MF-MLR model improved by only 4.6%. Among them, the accuracy of the MF-RFR model is higher than that of the original IMERG V06 product, with an obvious increase in dry days. The MF-RFR model-derived precipitation would lead to more accurate meteorological disaster assessments and hydrologic analyses than would otherwise be.

Hyperspectral anomaly detection (HAD) is a technique to find observations without prior knowledge, which is of particular interest as a branch of remote sensing object detection. However, the application of HAD is limited by various challenges, such as high-dimensional data, high intraclass variability, redundant information, and limited samples. To overcome these restrictions, we report an unsupervised strategy to implement HAD by dimensionality reduction (DR) and prior-based collaborative representation with adaptive global salient weight. The proposed framework includes three main steps. First, we select the most discriminating bands as the input hyperspectral images for subsequent processing in a DR manner. Then, we apply piecewise-smooth prior and local salient prior to collaborative representation to produce the initial detection map. Finally, to generate the final detection map, a global adaptive salient map is applied to the initial anomaly map to further highlight anomalies. Most importantly, the experimental results show that the proposed method outperforms alternative detectors on several datasets over different scenes. In particular, on the Gulfport dataset, the area under the curve value obtained by the proposed method is 0.9932, which is higher than the second-best method, convolutional neural network detector, by 0.0071.

The aim of this research is to analyze the influence of park design and location on their cooling effect during summer daytime in Barcelona. Spatial analytical methods, utilizing the land surface temperature data from the Landsat 8 satellite, were employed to assess the intensity and spatial extent of the Park Cool Island (PCI) in 86 parks over four consecutive years. The study investigated the influence of factors such as the proportion of green land cover, park shape, and surrounding characteristics on these PCI indicators using bivariate correlation and multiple linear regression analysis. Results revealed that 84 parks exhibited a positive cooling effect on their urban surroundings during the studied years, whereas the remaining parks showed no PCI. The statistical analysis indicated that these results are associated with the microclimatic uncertainty caused by the diverse relationships between the proportion of natural vegetated or pervious surfaces and the built spaces within the parks, proposed by the administration aiming to accommodate civic, cultural, and recreational usage. In conclusion, we discuss the parameters for modeling the interaction between park design and their urban surroundings in the context of climatic considerations. This information is crucial for leveraging urban parks as spaces that can mitigate the urban heat island effect while preserving their social value as community public spaces.

One of the Climate Absolute Radiance and Refractivity Observatory Pathfinder (CPF) mission’s science objectives is to intercalibrate the reflective solar bands of the NOAA-20 Visible Infrared Imaging Radiometer Suite (VIIRS) instrument against high-accuracy CPF measurements utilizing coincident, co-angled, and co-located footprints acquired over diverse Earth targets. To alleviate the effect of high polarization sensitivity of select VIIRS channels on intercalibration analysis, the CPF team will limit the intercalibration footprints over low-polarized scene types, which will be identified based on an empirical estimation of their degree of polarization (DOP) and angle of polarization (AOP) using Polarization and Directionality of the Earth’s Reflectance (POLDER) data. We describe the methodology for evaluating systematic errors in the estimation of DOP and AOP for Earth-reflected radiances using POLDER’s polarized bands and investigates their potential impact on CPF-VIIRS intercalibration uncertainty. The systematic errors were found to be <0.01 for DOP and <2.2 deg for AOP, which will have a negligible impact on CPF-VIIRS intercalibration uncertainty.

We aimed to combine the use of images obtained from remotely piloted aircraft systems (RPAS) and machine learning (ML) to identify the invasive alien species Psidium guajava in a protected area in southern Brazil. Field data were obtained in a sampling area, where the species’ geographic coordinates were collected with a global positioning system device. Remote data were collected with the Parrot Sequoia^® multispectral camera onboard the Phantom 4^® Pro platform. Image processing was used to generate reflectance maps and vegetation indices, after which four classes of interest were defined for model training. The supervised classification involved two approaches (pixel-based—BP and object-based image analysis—OBIA) and two ML algorithms compared (random forest—RF and support vector machine—SVM). For performance analysis, confusion matrices with user and producer accuracies, Kappa values and overall accuracy (OA) were calculated. The results demonstrate that the multispectral composition was excellent in identifying the invasive P. guajava, in an OBIA approach with the RF algorithm (0.90 Kappa and 93% OA). Thus, considering the priority of biodiversity conservation and the importance of the Brazilian Atlantic Forest for the maintenance of endemic and endangered species, we present a robust methodology to identify the invasive species P. guajava in subtropical forest, which can be applied in management strategies for the species control and eradication.

The traditional generative adversarial network (GAN) is widely used in the field of synthetic aperture radar (SAR) ground target image generation. However, GAN has the problem of unstable gradient update, which can easily cause the loss of image feature information, resulting in a low similarity between the generated image and the real image. To solve these problems, we propose an improved Wasserstein GAN with gradient penalty (IWGAN-GP), which introduces dense connection in the generator, integrates feature information at different levels to achieve feature reuse, and alleviates the gradient disappearance problem caused by deep networks. Moreover, by introducing the squeeze-and-excitation (SE) module into the densely connected network, on the basis of considering high-level semantic information and low-level geometric texture details, the optimal fusion weights of each channel can be automatically obtained to fully explore important target information in SAR images. IWGAN-GP alleviates the gradient disappearance caused by the depth of the network, strengthens feature propagation, and realizes feature reuse. It can automatically obtain the optimal fusion weight of each channel and improve the similarity between the generated image and the real image. The superiority of IWGAN-GP is verified on the datasets of MSTAR.

Climate change can have a serious impact on human life and occurs due to the emission of greenhouse gases, such as carbon dioxide, into the atmosphere. Deforestation is one contributing factor to climate change. It is important to understand the drivers of deforestation for mitigation efforts, but there is a lack of data-driven studies that can predict these drivers. In this work, we propose a contrastive learning architecture, called Multimodal SuperCon, to classify the drivers of deforestation in Indonesia using composite images obtained from Landsat 8. Multimodal SuperCon is an architecture that combines contrastive learning and multimodal fusion to handle the available Indonesian deforestation dataset. As a means to yield better performance, Multimodal SuperCon upgrades ordinary SuperCon so that it is able to handle auxiliary spatial variables as additional inputs. Our proposed model outperforms previous work on driver classification, giving an 8% improvement in accuracy compared to a state-of-the-art rotation equivariant model for the same task.

Uniform circular array electromagnetic vector sensor (UCA-EVS) can provide omni-directional and high-resolution azimuth information and polarization information, which has been widely concerned in the field of radar signal processing. Aiming at the joint estimation of direction of arrival (DOA) and polarization parameters of coherent sources via UCA-EVS, a multi-dimensional parameter estimation method with low complexity—polarization-DOA matrix method—is proposed. First, the rank of array covariance matrix is recovered by axial virtual translation, and then the direction and polarization parameters of the signal are estimated by using the eigenvalues and eigenvectors of the matrix based on constructing polarization-DOA matrix. Different from the traditional DOA matrix method, the proposed algorithm can not only estimate the azimuth information of the signals but also provide the polarization information of the targets, and the estimated parameters can be matched automatically. At the same time, it can estimate the parameters only by using the information of three elements, which can save hardware resources. In addition, the proposed method does not need to search for spectral peaks, which not only greatly reduces the computational complexity but also loses the estimation accuracy. Simulation results verify the feasibility of the proposed algorithm.

The registration of synthetic aperture radar (SAR) and optical images is a meaningful but challenging multimodal task. Due to the large radiometric differences between SAR and optical images, it is difficult to obtain discriminative features only by mining local features in the traditional Siamese convolutional networks. We propose a modality-shared attention network (MSA-Net) that introduces nonlocal attention (NLA) to the partially shared two-stream network to jointly exploit local and global features. First, a modality-specific feature learning module is designed to efficiently extract shallow modality-specific features from SAR and optical images. Subsequently, a modality-shared feature learning (MShFL) module is designed to extract deep modality-shared features. The local feature extraction module and the NLA module in MShFL extract deep local and global features to enrich feature representations. Furthermore, a triplet loss function with a cross-modality similarity constraint is constructed to learn modality-shared feature representations, thereby reducing nonlinear radiometric differences between the two modalities. The MSA-Net is trained on a public SAR and optical dataset and tested on five pairs of SAR and optical images. In the registration results of five pairs of test SAR and optical images, the matching rate of the MSA-Net is 5% to 15% higher than that of other compared methods, and the matching errors of the matched inliers are on average reduced by about 0.28. Several ablation experiments verify the effectiveness of the partially shared network structure, the MShFL module, and the cross-modality similarity constraint.

Most of the existing methods for indoor point cloud semantic segmentation employ sophisticated structures to obtain local features, leading to high complexity. In addition, the current prevailing networks focus less on the extraction of global features of point clouds, which may damage the performance of the networks. Moreover, there are few networks that achieve the balance between accuracy and efficiency. To solve the above problems, we propose an improved network architecture based on anisotropic separable set abstraction network. It is more accurate and more efficient. First, we introduce an improved ASSA module to comprehensively consider the influence of the distance between the neighbor and the centroid on the neighborhood features, to better obtain the local features of the point clouds. Later, we design an inverse residual module, which optimizes the backbone network by scaling the receptive field to improve the extraction ability of the global features of the point clouds. Then we use a mixed pooling method to solve the disorder of point clouds and fuse both the coarse-grained and fine-grained features. Evaluated on the stanford large-scale 3D indoor spaces dataset, the experiment results illustrate that our method not only improves the extraction of global features and local features effectively, but also achieves an advanced performance among several mainstream point cloud semantic segmentation methods. Overall accuracy is 89.7%, mean class accuracy is 75.6%, mean intersection over union is 69.5%, and floating point operations are 15.9G.

Dimensionality reduction techniques can remove redundant information from hyperspectral images (HSIs) and improve discriminability. However, due to the inherent nonlinear characteristics of HSI, there may be non-Euclidean structures in the data and its topological properties may make it suboptimal to recover the low-dimensional manifolds by means of a linear projection. As a result, linear projection from high-dimensional space to low-dimensional discriminative space is not always effective. To better explore the intrinsic geometric structure of HSI, we propose a Gaussian manifold metric learning (GMML) method, which performs explicit and nonlinear dimensionality reduction for HSI. It models pixel neighborhoods as Gaussian distributions to combine spatial and spectral information. In GMML, samples are mapped as elements on the Gaussian manifold consisting of Gaussian distributions to retain the intrinsic characteristics of data. By treating the inner product of the tangent space on the manifold as a kernel function, our method performs metric learning in the tangent space. Therefore, the proposed GMML method makes intra-class samples more compact and inter-class samples more separated while preserving the geometric structure information inherent in the data. Experiments on six real hyperspectral datasets validate the effectiveness of the proposed GMML algorithm to extract discriminating features compared to several related methods.

Pansharpened images are frequently utilized as base data in image classification applications. Nonetheless, the accuracy of image classification heavily relies on the efficiency of the pansharpening strategy applied. With numerous existing pansharpening approaches available, it becomes challenging for analysts to select the one that yields the best outcome. Recently, the deep learning (DL)-based pansharpening approaches have become popular due to their capabilities. Thus, this study aims at examining the image classification performance of pansharpened images generated by several commonly used DL-based pansharpening algorithms that rely on pre-trained models and comparing them with those of several traditional pansharpening algorithms. The experiments that were conducted in two test sites indicated that the DL-based pansharpening algorithms could be used in various circumstances. It can also be inferred that the DL-based pansharpening algorithms enhanced the image classification accuracy more effectively than many other traditional algorithms. The use of pre-trained models led to robust pansharpening results and, therefore, superior classification accuracies in most cases.

Image dehazing is an important preprocessing task since haze extremely degrades the image quality and hampers the application of remote sensing vision system. Although the deep learning-based method has been successful in image dehazing, there has been little effort to harmonize convolutional neural networks and transformer to better satisfy removing haze. In particular, local and global representation learning are equally important for the challenging image dehazing task. To this end, we propose an effective ensemble dehazing network (EDHN) for visible remote sensing images. Specifically, we introduce two key backbone modules for the developed ensemble framework, including operation-wise attention module and transformer module. The operation-wise attention module is designed for restoring spatially varying degradation, and the transformer module is employed to refine haze-free background textures and structures. Furthermore, residue channel prior and feature aggregation block are also incorporated into our ensemble architecture to further guide image reconstruction and boost image restoration. Experimental results show the superiority of our proposed EDHN and demonstrate the favorable performance against recent dehazing approaches.

Unsupervised feature extraction techniques of hyperspectral images (HSIs) have recently drawn significant attention for their excellent performance and efficiency in classification. In some existing methods, the denoising process that reduces the influence of inherent noise is ignored, and the nonlinear edge characteristics and multi-scale features that help to classify still need to be fully considered. To solve these issues, we employ a multi-scale nonlinear edge-based unsupervised three-phase model (UTPM) for hyperspectral feature extraction. Specifically, in the initial phase, a noise-adjusted principal components technique is adopted to lower the noise to improve the performance of the proposed model. Then, a neighbor band grouping technique is designed to reduce redundancy and computational cost with information entropy. Because the information entropy can concretely reflect the importance of different bands in the same group, the inner structure can be maximally preserved. Finally, we utilize a multi-scale feature fusion on kernel low-rank entropic analysis to extract nonlinear edge features and combine it with a convolution algorithm to fuse the elements of multiple scales to improve the classification performance. Compared with several other classical or progressive unsupervised hyperspectral feature extraction algorithms, the classification results on three public HSI datasets validate the effectiveness of UTPM.

Automatic registration of medium-resolution remote sensing images of coral reefs, particularly those without artificial facilities, faces two challenges: difficulty in identifying the same coral reefs in different images and instability of the fine-tuning process. To overcome these challenges, we propose an automatic registration method that combines morphological information pairing with constrained iterative fining. This method comprises three steps. First, the contours of the coral reefs were extracted using level set method. Subsequently, the same coral reefs in the two images were identified and paired based on morphological similarities and relative locations. Finally, iterative fine registration with a constrained strategy was performed by controlling abnormal changes in the geometric center to further improve the registration accuracy for every pair of coral reefs. The proposed automatic registration method was validated using a Landsat5 image acquired on April 15, 2005 and a HJ-1B image acquired on May 4, 2010. Compared with the scale-invariant feature transform (SIFT) method and the SIFT with Random Sample Consensus method, the proposed method showed good performance in the automatic registration of coral reef images.

Since its launch in May, 2002, Aqua MODIS has successfully operated for more than 20 years and has continuously generated a wide range of data products that have enabled and supported the remote sensing community and users worldwide for their studies of the Earth’s system by monitoring changes in its key environmental parameters. Although Aqua MODIS, designed with a lifetime requirement of 6 years, is currently operated in its extended mission phase, it continues to make high quality global observations of the Earth’s surface via its 36 spectral bands that cover wavelengths from visible to long-wave infrared. To date, all instrument on-board calibrators (OBC) remain capable of performing their design functions, providing various calibration data sets to help monitor on-orbit changes in sensor responses and performance characteristics. In addition to the OBC, regularly scheduled lunar observations and select Earth-view targets are used extensively to support sensor on-orbit calibration, especially for the calibration of the visible channels (or bands). We provide an overview of Aqua MODIS on-orbit calibration activities and methodologies for both reflective solar bands and thermal emissive bands, illustrate its on-orbit performance over the past 20 years using examples derived from OBC measurements, lunar observations, and Earth-view response trends, and describe various calibration improvements made over its entire mission. We focus on key issues identified since launch, such as solar diffuser degradation, electronic crosstalk, and on-orbit changes in sensor response versus scan-angle, along with approaches and strategies developed to mitigate their impact on sensor calibration quality. Also discussed in this paper are some of the key calibration enhancements incorporated recently in the Collection 6.1 and the upcoming Collection 7 Level-1B algorithms.

Providing unbiased ground truths for large size images is a complex task that is difficult to achieve in practice. The aim of our method is to easily produce reliable ground truths, from images covering large areas while respecting the physical nature of the observed data. The first step localizes all classes existing in a large size image without any a prior knowledge, as well as the samples that make them up. Next, the user selects classes from these unbiased detected classes, in order to build a true ground truth adapted to this application. No transformation is carried out on of characteristics (spectral features) of the hyperspectral image pixels measured objectively by the sensor. To prove the relevance of the learning sample selection method, we use three supervised classification methods requiring training samples. Two hyperspectral images are selected to illustrate the performance of the obtained training samples. The results of the proposed method are also compared to those available of some state of the art deep learning methods.

We propose a pioneering approach for gathering data on the forest canopy, one that merges two cutting-edge technologies: ConvNext tiny and simple linear iterative clustering (SLIC) (ConvNeXt and SLIC vectorize for tree species mapping, CSVTSM). By leveraging the clustering label generated by SLIC, CSVTSM obtains the vectorized result of land cover and allows us to obtain the location and distribution range information about the individual canopy. Moreover, CSVTSM employed ConvNeXt tiny, an advanced model, to identify vector objects and produce vector tree species maps across the experimental area to obtain the area of various tree species. To evaluate the accuracy of our methodology, we compared the canopy boundary of the original image to the four experimental areas that were vectorized by CSVTSM. Our model’s efficacy was assessed by calculating the overall accuracy and Kappa coefficient on the validation set. The effectiveness of the proposed method was qualitatively evaluated by calculating the difference between the tree species area extracted by CSVTSM and manual extraction. The results indicate that, when considering the same canopy size, the vectorization approach based on the SLIC-based clustering label generates vector boundaries that are more closely aligned with the distribution of actual canopy boundaries in the experimental area. Furthermore, when used in conjunction with the SLIC vectorization approach, the ConvNeXt model can produce an even more precise vector map of tree species and more accurate tree species area information is obtained (only a 0.26 ha difference from manual extraction). These findings demonstrate that CSVTSM can be used to quantitatively acquire a wealth of information on individual tree positions, crown distribution ranges, and planting areas from high-resolution RGB photos captured by low-consumption unmanned aerial vehicle platforms. The implications of these results are wide-ranging, with local managers expressing keen interest in our findings as they offer trustworthy support for statistical application fields, such as gathering area ecological information, locating tree species resources, and disseminating spatial information about forests..

Accurate crop type and crop growth stage maps are essential for agricultural monitoring and ensuring food security. A wide variety of airborne and spaceborne sensors now provide high spatial, spectral, and temporal resolution images, which are vital for crop mapping and monitoring. Crop type and its growth stage can be characterized by spectral, spatial, and temporal features. The classification of crop types and growth stages has been explored in previous studies as independent tasks. However, the growth stages of a crop are an important factor in identifying the crop and vice-versa. A multi-task learning (MTL) framework is proposed in this work to classify the crop type and its growth stages simultaneously. A hybrid convolutional neural network and temporal convolutional network (CNN-TCN) architecture is presented to process a multitude of features relevant to the tasks. To learn the spatio–spectral features, we fed the hyperspectral input to 3D convolution blocks and multispectral input was given into 2D convolution blocks. We reformulate these multi-channel features into two dimensions and feed them into the temporal convolutional neural network. Subsequently, we use two fully connected branches for each task. MTL frameworks were developed for multispectral (Mx), hyperspectral (Hx), and the combination of Hx and Mx (Hx-Mx) images to model crop type and crop growth stage classification. Results reveal that the proposed model for Hx-Mx outperformed the best single-task model by 13% and 8% in crop growth stage and crop type classification, respectively. Compared to single-task models, the proposed model can exploit the high spectral information from Hx images and high spatial information from Mx images, making the proposed model more useful for unmanned-aerial-vehicle-based crop mapping.

Electro-optical and infrared propagation within the troposphere is strongly determined by the atmospheric path along the line-of-sight. The industry standard program moderate resolution transmittance (MODTRAN) model can be used to approximate path transmittance and radiance from the ultraviolet to the infrared. To understand the variability in the path effects (transmittance/radiance), one must accurately represent the environment in which they are operating. One significant knowledge gap is a universal and consistent means to inform the aerosol component of the atmospheric modeling within MODTRAN. The aerosol component can be highly variable, which can result in significant changes to the overall path transmittance and radiance depending on aerosol size and type. A fusion of European Centre for Medium Range Weather Forecasts Re-Analysis 5 atmospheric profile data and statistical draws from the global NOAA-ISD visibility archives are presented to construct regionally specific atmospheric inputs to MODTRAN. Using a 6 km horizontal path geometry, a year’s worth of hourly (8760 times) data is modeled to determine the path transmittance. The results are aggregated via empirical cumulative distribution functions of the path transmittance. We determine the type of day one is operating within by defining a bad day as the 10th percentile low transmission, an average day as the 50th percentile, and a good day as the 90th percentile high transmission. Using the percentile atmospheres determined by transmittance, one can analyze expected performance levels from current sensor systems, e.g., if a detection threshold is met on the 50th percentile day, one can expect this threshold to be met 50% of the time. In addition, the specific atmospheres (visibility, pressure, temperature, humidity, rain, etc.), which were determined to result in the 10th, 50th, or 90th percentile day, can be analyzed to build synthetic scenes, construct modeling environments, and ultimately drive design and system trades to optimize performance from the visible to the longwave infrared.

In recent decades, remote sensing of vegetation by hyperspectral imaging has been of great interest. An important part in interpreting the remotely sensed spectral data is played by simulators, which approximate the connection between plants’ biophysical and biochemical properties and detected spectral response. We introduce improvements and new features to recently published hyperspectral leaf model HyperBlend. We present two methods for increasing simulation speed of the model up to 200 times faster with slight decrease in simulation accuracy. We integrate the well-known PROSPECT leaf model into HyperBlend allowing us to use the PROSPECT parametrization for leaf simulation. For the first time, we show that HyperBlend generalizes well and can be used to accurately simulate a wide variety of plant leaf spectra. HyperBlend is available as an open-source Python project under MIT license in a GitHub repository available at: https://github.com/silmae/hyperblend.

Ground point identification and segmentation are fundamental to the light detection and ranging (LiDAR) based environment perception because they affect the accuracy and computational efficiency in the following data processing steps. A common problem that results in over- and under-segmentation occurs when the objects of interest are nonhomogeneous, and the sampling density is uneven. This study addresses this issue using grid- and homogeneity-based approaches. This work began with a combined conditional and voxel filtering approach to shrink the spatial range and reduce the amount of point-cloud data. The spatial range was then divided into a concentric circular grid to reduce the complexity of data processing. A dynamic threshold model was used to classify the cloud points to improve the accuracy of ground-point identification on uneven, broken, and sloped roads. Additionally, a point cloud homogeneity model was used to optimize the ground point identification results in areas with vegetation. The experimental study was conducted based on the data provided in the semantic KITTI dataset, wherein comprehensive comparisons were made with state-of-the-art algorithms. The average precision, recall, F1, and running time of the proposed method were 92.5%, 90.89%, 0.92, and 0.146 s, respectively, outperforming most of the selected models in balanced accuracy and computational efficiency.